Today's Highlights
· Nature sub-journal: a single grasping model drives 7 different robotic hands zero-shot, with 94.3%–98.0% success rates
· HKU-led RoboDojo unified benchmark: the strongest model achieves only 12.8% real-robot success, versus 100% for human experts
· WeRide's H1 revenue reached RMB 350 million, up 73% YoY, with overseas/international revenue up 154%
· Waymo has imported over 3,200 Zeekr-chassis Robotaxis, with per-vehicle cost estimated at roughly half the previous generation
· A3: North American Q2 robot orders reached 8,940 units, while automotive OEM orders fell 25% in H1
· FCC revokes drone certification for a "U.S.-branded" front company for the first time — Odyssey Robot loses approval
· Zhongke Guiji (Chinese tactile-sensor startup) places a single order for 200,000 tactile sensors, enough for 40,000 dexterous hands
· Ukraine has ordered over 3,000 ground robots cumulatively, with 640 units in July alone
I. Research Progress
One grasping model works across seven robotic hands — SpaHybGen published in Nature Machine Intelligence · manipulation
Robotic grasping has long been stuck in a dilemma: deep-learning policies are sample-efficient but locked to specific hardware, while analytical planners transfer across hands but struggle with sensing noise. SpaHybGen bridges the two — first training a neural network to infer a hardware-agnostic "spatial contact feature" representation directly from noisy depth observations, then handing it to a differentiable optimizer that solves for stable grasps on the specific hand geometry. Results from a single training run zero-shot drove seven different hands, from two-finger to five-finger grippers, with no hardware-specific retraining at all, achieving 94.3%–98.0% grasp success in semi-cluttered scenes and supporting dynamic 20 Hz grasping and multi-hand coordination in dense, cluttered environments. The authors decouple perception from action around a shared contact interface, and have open-sourced the code, model, and the generated contact-feature dataset.
Nature Machine Intelligence (2026) · Code and model open-sourced on GitHub (SpaHybGen) · Coverage: Nature source
GeniWorld: helping world models "understand" robot actions — π0 real-robot success rate rises from 40.8% to 69.0% · world-model
Expensive real-robot data and hard-to-build scenes remain long-standing bottlenecks for generalizing embodied policies. GeniWorld, jointly proposed by the Tsinghua Shenzhen International Graduate School, Tencent Robotics X, HKUST, and Shenzhen Technology University, uses URDF rendering to convert numerical action sequences into a "visualized action representation" aligned with real camera viewpoints, then concatenates it with noisy video latents and feeds them into a causal diffusion Transformer to predict the future — letting the world model be precisely controlled by actions rather than merely generating passively. On the real-robot side, using a dual-arm Xtrainer platform with only 25 demonstrations per task to train π0, adding GeniWorld synthetic data raised the overall success rate across four tasks (moving a bowl, folding a towel, etc.) from 40.8% to 69.0%, with spatial-rearrangement scenarios rising from 37.5% to 70.0%.
Tsinghua Shenzhen International Graduate School × Tencent Robotics X × HKUST × Shenzhen Technology University · Coverage: Sina Finance source
Flex-π: a frozen video VAE turns out to encode 3D point maps too, for free · world-model
Nearly all world-action models (WAMs) predict only RGB latents and train purely on pixel-level reconstruction, with no explicit supervision on the 3D geometry and object semantics that manipulation actually needs. The authors found a "free lunch": the frozen video-generation VAE used to encode RGB also encodes 3D point maps almost losslessly, with no point-map-specific training required. Building on this, the 6B-parameter model projects geometry, DINO semantics, and RGB into the same latent space within a Mixture-of-Transformers backbone, denoising them jointly with actions, while per-stream dropout lets a single checkpoint run on any subset of streams. On real dual-arm fine manipulation tasks, it outperforms the strongest baseline by 2–7x at best, both in- and out-of-distribution, and infers faster than π0.5.
Ge Yan et al. · arXiv 2608.10860 source
FACT: making failed actions usable as training targets too · world-model
arXiv has seen a cluster of world-action model papers today, and FACT targets a gap most of them skip: these models are almost always trained only on successful demonstrations, so they have no reason to learn what "bad actions lead to," making them prone to hallucinating successful futures during inference. FACT instead predicts future video and task progress conditioned on the actions actually executed, turning failed rollouts into legitimate supervision signals rather than discarded data; the progress predictor sees both successful and failed outcomes, and can score candidate actions at inference time. The authors report continued improvement as failure data is added in both simulation and real dual-arm experiments, with a significant reduction in success-biased hallucination under bad actions.
Quanquan Peng et al. (including Nicklas Hansen, Xiaolong Wang) · arXiv 2608.10232 source
SALT: action tokenizers optimize only for reconstruction, losing the semantics of verbs · vla
Action words describe not just outcomes but "how" something is done, yet VLA action representations are typically optimized in raw action space using L1/L2 reconstruction loss — numerical closeness doesn't equal linguistic meaning. The authors first show on BridgeV2 that action trajectories carry verb-grounding information beyond visual state changes, and that reconstruction-only discrete tokenization systematically erases it; SALT then adds an auxiliary objective to VQ-VAE-style tokenizers, requiring a frozen vision-language model to recover the episode's instruction from the quantized action latent. Policies trained with SALT achieve an average success rate of 71.9% on SimplerEnv, versus 42.7% for a reconstruction-only VQ-VAE tokenizer and 31.2% for FAST.
Li Wenjie et al. · arXiv 2608.10484 source
4D-WAM: driving world models shouldn't learn only 2D projections · autonomy
Existing driving world-action models are mostly trained on video, but video is only a 2D projection of the underlying 4D scene. 4D-WAM uses geometric foundation models to provide training-time supervision, modeling scene evolution in a 4D-consistent space, achieving state-of-the-art results on both the NAVSIM-v1 and NAVSIM-v2 benchmarks.
Jiacheng Fu et al. · arXiv 2608.10107 source
Hidden in Plain Sight: "invisible" attacks on VLAs using diffusion models · vla
Adversarial robustness for VLAs has so far received little systematic study, yet when these models are misled the consequences land directly in the physical world. The authors use diffusion models to construct attack samples unconstrained by perturbation norms — looking like natural scenes rather than noisy images — to probe VLA failure boundaries on real manipulation tasks. Worth watching closely as embodied models move faster into production lines.
Jiahui Han et al. · arXiv 2608.10393 source
Other papers today: JEPA-WAM does stage-level joint embedding prediction using a frozen V-JEPA2 encoder (arXiv 2608.10780 source); Surgical WAM does data-efficient surgical robot learning based on Cosmos Policy (arXiv 2608.11204 source); full-body planning for humanoid robots in confined spaces, training a residual RL policy with self-collision avoidance as reference (arXiv 2608.10220 source); single-view dual-arm coordinated dexterous grasping of large objects (arXiv 2608.10383 source); VIScore diagnoses the "planability" of latent-space world models (arXiv 2608.11174 source); HUI360, currently the largest 360° first-person dataset for human-robot interaction anticipation (arXiv 2608.11051 source); Gated VLA-Cache uses neural introspection to decide whether to reuse the KV cache (arXiv 2608.10824 source); Dreamer-SAC puts off-policy SAC directly inside a latent-space world model for autonomous driving (arXiv 2608.10386 source); DriveVLA-M0 builds a failure-case memory bank for retrieval augmentation (arXiv 2608.10413 source); AECNav does training-free zero-shot open-vocabulary object navigation via evidence gating (arXiv 2608.10817 source).
Open Source · Tools · Benchmarks
· RoboDojo: a unified manipulation benchmark led by the University of Hong Kong's MMLab together with nearly 20 universities (including UC Berkeley and Tsinghua), bringing simulation evaluation, standardized real-robot testing, and policy benchmarks into a single framework covering 42 simulation tasks, 18 real-robot tasks, and 30 representative policies, assessing generalization, memory, precision, and long-horizon execution. The first results deliver a sobering number: the best-performing model reaches 8.80% success in simulation and 12.8% on real robots, versus 76.03% and 100% for human experts respectively. In its first week, the project drew over 100,000 views on X, and its open-source resources have been downloaded over 100,000 times on Hugging Face. source
· NVIDIA JetPack 7.2.1: adds Agentic Video Skills and T3000 simulation support. source
· LeRobot Studio / ROSView: an open-source offline visualization tool for embodied data from IO-AI Tech (Chinese embodied-data infrastructure startup); the same release also included several new real-world datasets. source
II. Funding and Deals
Westlake Robotics | Series A | over RMB 100 million (RMB 500 million across four rounds in six months) · embodied
Investors include Sino-French Fund (Saif), Xiaomiao Chengcheng Capital, and Haiyuan Fund. Founded in 2024 and headquartered in Hangzhou, the company is Westlake University's first commercialization project in AI and robotics, pursuing a full-stack approach of "general-purpose brain + humanoid full-body cerebellum + proprietary humanoid platform," with plans to launch its fully self-developed humanoid, Westlake o1, in 2026. Founder Wang Donglin is a tenured professor at Westlake University and deputy director of its AI department; co-founder Zhang Yue is likewise a tenured professor there. The company says it holds nearly RMB 100 million in orders, spanning STEM education training, humanoid data collection, and power-grid inspection, with new funds focused on developing a unified full-body humanoid foundation model. Source: Shanghai Securities News · China Securities Journal (via Eastmoney) source
Demak Precision | Series C | RMB 420 million · hardware
Led by Furong Investment, with participation from Changxing Industrial Group, Nanhu Equity Fund, Yuanjing Capital, and Deqing Industrial Investment, plus a follow-on from existing investor Yunqi Capital. Founded in 2016 in Changxing, Huzhou, Zhejiang, the company's core business is precision machining of semiconductor-equipment chambers and large structural components for solar and wind power, with nearly a thousand imported high-end machine tools and over 20,000 square meters of temperature-controlled factory space. It is now applying that batch-consistency capability to humanoid robot torsos, joint housings, and limb connectors, having already established partnerships with Unitree Robotics, Zhiyuan Robotics (Chinese humanoid startup), and UBTech. As production scales from hundreds to tens of thousands of units, the yield bottleneck shifts to accumulated tolerances in structural components rather than algorithms — this round, with heavy participation from local state capital, directs funds primarily toward physical manufacturing capacity. Source: Robot Lecture Hall source
IO-AI Tech | new round | several hundred million RMB · adjacent
Jointly invested by Shunwei Capital, Matrix Partners China (Songhe Capital), and Shenzhen Capital Group, with strategic investment from a leading robot-platform company. Founded in 2023 in Shenzhen, the company builds embodied-data infrastructure rather than pure data-collection outsourcing, with a product line covering the general-purpose teleoperation collection system TeleXperience, the real-world human data collection system SenseXperience, and the data management platform EmbodiFlow. The company says it has served over 150 robot-platform and model companies worldwide, adapting to over 80 robot models cumulatively. Source: PEdaily source
Qiaojie Digital Matter | Pre-A+++ round | over RMB 100 million · embodied
Led by China Mobile Chain-Leader Fund, with follow-on investment from existing investors Fosun Chuangfu and Shenzhen Capital Group. The company builds a general-purpose robot "cerebellum," focused on bipedal and quadrupedal locomotion control, with its solution already covering over 20 robot models; it says it can compress the traditional 3–4 month locomotion-control development cycle down to weeks, with algorithms iterating monthly. The funds will support the strategic rollout of its "general-purpose robot operating system." Source: Guandian.cn source
Alloy Robotics (San Francisco, USA) | seed round | $8 million | valuation approx. $80 million · adjacent
Led by Square Peg. The product is an AI agent specialized in root-causing robot failures — CEO Joe Harris's observation is that the answer usually already sits in the robot's own logged data, but engineers must dig through it one run at a time across manually shared files, cloud storage buckets, SQL queries, Python scripts, and replay tools, taking hours to days. The company says its customer count has grown over 50% month-over-month since commercialization, and it has processed data from over 10,000 robot runs. Source: Forbes source
Cambridge Aerospace | new round | $300 million · adjacent
A defense drone developer; this round is among the largest in the current defense autonomous systems sector. Source: The Robot Report source
Turing (Japan) | planning U.S. office | targeting $10 billion IPO valuation · autonomy ⚠️ Company projection
The Japanese autonomous-driving startup plans to establish a U.S. office and has set a target IPO valuation of $10 billion. Both the target valuation and listing timeline are the company's own plans, with no offering arrangements in place yet. Source: Guandian.cn source
III. Commercialization and Deployment
WeRide's H1 revenue reaches RMB 350 million, up 73% YoY, with overseas/international revenue up 154% · autonomy
WeRide (NASDAQ: WRD, HKEX: 0800) reported Q2 and H1 2026 results: total H1 revenue of RMB 350 million, up 73% year-over-year; Q2 revenue of RMB 230 million, up 82% year-over-year and up 103% quarter-over-quarter, with gross margin rising to 37.5%, 9.4 percentage points higher than the same period in 2025. Overseas/international expansion is the main driver of growth, with H1 overseas/international revenue up 154% year-over-year and Q2 up 164% year-over-year. As of July 31, 2026, the company's global L4 fleet numbered approximately 3,400 vehicles, including over 1,800 Robotaxis; Q2 L4 business revenue rose 131% quarter-over-quarter, and Robotaxi registered users rose 35% quarter-over-quarter. The other growth leg is mass-production ADAS: L2++/L3 business revenue rose 219% quarter-over-quarter, with roughly 30,000 units of its self-developed end-to-end system WRD 3.0 shipped, production designations secured for over 30 vehicle models, and a joint L3 feature proof-of-concept underway with a German luxury automaker. The business now covers 13 countries and over 60 cities. Source: SmartHey source, Bloomberg source
Waymo has imported over 3,200 Zeekr-chassis Robotaxis, with per-vehicle hardware cost estimated roughly 48.5% below the previous generation · autonomy ⚠️ Estimate
According to Forbes reporting, corroborated by customs records, Waymo has cumulatively imported over 3,200 Ojai Robotaxis built on Zeekr's platform, with more than 2,600 recorded in 2026 alone; over 500 were present simultaneously at its Mesa, Arizona vehicle-integration facility in August. These vehicles are manufactured in Ningbo by Geely-owned Zeekr and imported without autonomous-driving systems installed — the Chinese-made portion covers the body, battery, and drivetrain, with Waymo's sixth-generation Driver (4 lidars, 6 radars, 13 cameras) fitted in the U.S. On cost, U.S. customs filings put the CM1e chassis factory price at approximately $38,000–$38,500; applying the upper bound plus a 102.5% tariff puts the landed cost at approximately $78,000. Adding Waymo's disclosed sixth-generation hardware cost of approximately $25,000 brings the total vehicle-plus-autonomous-hardware cost to roughly $103,000 — about 48.5% below the previous-generation Jaguar I-Pace setup, estimated at approximately $200,000 (vehicle $75,000 plus autonomous hardware $125,000). It should be noted that final integration cost is not publicly disclosed, so the $103,000 figure does not equal total landed vehicle cost; this comparison is based solely on the vehicle and hardware figures that have been disclosed. Source: CarNewsChina source
Ukrainian military has ordered over 3,000 ground robots cumulatively, with 640 units in July alone · adjacent
Via the military-technology marketplace operated by the state-backed innovation platform Brave1, the Ukrainian military has ordered over 3,000 ground robots since the marketplace launched last year, with over 640 units purchased in July alone — a 680% increase over the 2025 monthly average. Brave1 Market product manager Yuliia Myrna says ground robots and interceptor drones are the two fastest-growing categories in 2026; a single ground robot can cost tens of thousands of dollars, more than most drones, but offers diverse functionality and longer service life. Ukraine is deliberately accelerating their use for high-risk tasks such as logistics and casualty evacuation, incentivizing troop adoption through a points-based system. Brave1 CEO Andrii Hrytseniuk says the number of companies engaged in ground-robot production has grown from a handful before the war to about 280 today. Source: SSBCrack source
Unitree's cumulative humanoid robot production reaches approximately 18,000 units · humanoid
Following this week's IPO subscription being oversubscribed 8,288 times, Unitree Robotics disclosed that its cumulative humanoid robot production has reached approximately 18,000 units. This figure refers to cumulative units off the production line, not deliveries or shipments, and cannot be directly compared with recent industry shipment-ranking figures. Source: Jiemian News source
AgiBot livestreams sorting at 1,816 items/hour, with simpler hardware · embodied ⚠️ Company claim
AgiBot Robotics (Chinese embodied-AI startup, formerly known as Zhiyuan Robotics) held a one-hour fully autonomous logistics-sorting livestream, continuously sorting real packages of random size, shape, material, and weight at a speed exceeding 1,816 items/hour with over 98% accuracy, which the company says is about 45% faster than the 1,248 items/hour previously demonstrated by a U.S. company. Notably, the hardware direction is the opposite of what one might expect: AgiBot uses dual robotic arms with grippers rather than a "humanoid plus five-fingered dexterous hand" setup — partner and algorithm lead Gan Ruyi says this cuts hardware cost by 70%, while placing higher demands on the embodied foundation model, backed by their unified world model WALL-B, which fuses vision, language, touch, action, and physical prediction. A one-hour livestream demonstrates capability but is still a separate question from stability under continuous production-line operation. Source: Sohu (reposted) source
IV. Industry Developments
A3: North American Q2 robot orders reach 8,940 units, while automotive OEM orders fall 25% in H1 · industrial
According to data from the Association for Advancing Automation (A3), North American companies ordered 8,940 robots in Q2 2026, worth $622 million — up 4.3% year-over-year in units and 21.3% in value; H1 totals reached 17,995 units worth over $1.16 billion, up 2.0% and 6.6% respectively versus the same period in 2025. The real story is structural: H1 automotive OEM orders fell 25%, while semiconductor and electronics orders grew 35% and life sciences/pharmaceuticals grew 32%, exactly offsetting the automotive weakness. In Q2, semiconductor and electronics orders were up 38% year-over-year, automotive parts up 20%, and food/consumer goods and metals each up 18%. Non-automotive customers accounted for 56% of Q2 order volume, continuing the trend of robots spreading into more industries. Collaborative robots accounted for 2,774 units and $114 million in H1 orders — 15.4% of units and 9.8% of value — with especially strong penetration in life sciences and semiconductors, at 43.7% and 36.5% of each sector's H1 orders respectively. A3 executive vice president Alex Shikany says automotive remains an important source of demand, but the breadth of growth outside automotive OEMs is a trend worth watching closely. Source: DC Velocity source
FCC revokes equipment certification for a "U.S.-branded front" company for the first time · adjacent
The U.S. Federal Communications Commission issued a revocation order (DA 26-839) on August 11, permanently withdrawing drone and controller equipment certification from Delaware-registered Odyssey Robot LLC. Under rules effective December 2025, drones and key components manufactured in covered foreign countries are placed on a "Covered List," barring them from the equipment certification required for import, marketing, and sale. The FCC's investigation found that Odyssey claimed its products were designed in California and assembled by eTak Worldwide in Grand Prairie, Texas, but found no evidence of any domestic-manufacturing records — constituting a false statement. The investigation was triggered by research published in June 2026 by independent security researcher Konrad Iturbe; a prior order to show cause gave Odyssey a 10-calendar-day window to respond, which it did not use. This marks the FCC's first use of its expanded national-security authority to retroactively strip equipment approval from a company alleged to be a "U.S. front" for foreign-made hardware, with chairman Brendan Carr previously stating he is monitoring other white-label import brands using similar tactics. The FCC is also weighing whether to broaden a wider "military-grade" definition to include lidar, thermal imaging, and spraying payloads, expanding the scope of restrictions. Source: The Drone Girl source
Honor unveils Robot Phone, packing dexterous-hand-grade joint motors into a smartphone · hardware ⚠️ Company claim
Honor held a global launch event in Guangzhou, unveiling the Robot Phone, priced at RMB 9,999 (12GB+512GB) and RMB 12,999 (16GB+1TB), going on sale August 18. What's more interesting for Physical AI readers than the phone itself is its four-degree-of-freedom titanium-alloy dexterous gimbal: the gimbal motor weighs just 2.6 g, which Honor says is 34% smaller than dexterous-hand joint motors used in embodied AI, with machining precision reaching ±0.005 mm; the mechanical system integrates over 100 precision components, with a maximum tri-axis control speed of 360°/s. On the software side, the system is powered by an Agentic OS kernel, debuting the YOYO Pro mode, co-developed with Alibaba's Qwen for on-device capabilities; the gimbal can automatically frame shots following hand gestures, music rhythm, and call participants. Claims such as "world's first" and "pioneering" are the company's own framing, but once consumer-electronics production scale genuinely drives down the cost of this kind of miniature high-precision joint module, the benefits could extend well beyond smartphones. Source: Phoenix.com Reviews source
Dyna Robotics releases world-action model Dyna-2, claiming the first validation of a cross-embodiment scaling law · world-model ⚠️ Company claim
Dyna Robotics has released Dyna-2, a single generative model capable of denoising future video and actions either jointly or separately, pretrained on a dataset of over 1 million hours of first-person human manipulation video, 43.8 million clips, 97,160 task instructions, and 9,917 object types. The company says this is the first time a scaling law has been validated across platforms with differing embodiment — as pretraining data scales from 1,000 hours up to 1 million hours, the model's predictive performance on held-out human data rises monotonically, with an inflection point between 10,000 and 100,000 hours. Compared under matched conditions with the company's previous-generation VLA architecture, Dyna-1, the world-action model achieves a success rate 1.55x that of the VLA; more telling are the deployment numbers: both generations achieve 100% success in internal testing, but in actual customer environments Dyna-1's pass rate was 46% versus 87% for Dyna-2. The company was founded in 2024, is headquartered in the U.S. with a hardware R&D center in Shanghai, and closed a $120 million funding round last September with participation from Nvidia, Samsung, LG, and Amazon. All figures above are the company's own claims, not yet independently reproduced. Source: iFeng/Zhidongxi source
Shanghai Electric and Shanghai FANUC Robotics form joint venture Shangfa Technology, registered capital RMB 200 million · industrial
Shangfa Technology (Shanghai) Co., Ltd. was recently established, with legal representative Qian Hui and registered capital of RMB 200 million; its business scope includes intelligent robotics R&D, industrial robot manufacturing and installation/maintenance, and AI infrastructure software development. Equity records show the company is jointly held by Shanghai Electric (601727.SH)'s wholly owned subsidiary Shanghai Electric Automation Group and Shanghai FANUC Robotics. Source: Sina News (citing Qichacha) source
Einride partners with DAF Trucks to integrate the Einride Driver into production trucks · autonomy
Dutch truck manufacturer DAF will integrate Einride's autonomous driving system, Einride Driver, to advance autonomous electric freight at scale. Having the OEM itself, rather than a retrofitter, serve as the integrator is the typical path for driverless trucking moving from demonstration fleets to mass-production installation. Source: The Robot Report source
Pudu unveils ET1, bringing commercial floor-scrubbing robots into 100–800 square meter small stores · adjacent
Pudu Robotics has launched the PUDU ET1, the first product in its new E-series, targeting small commercial spaces with cleaning areas of 100–800 square meters, such as convenience stores, pharmacies, chain restaurants, budget hotels, and offices. The product combines floor scrubbing, sweeping, vacuuming, and mopping into a single pass — sweeping in front, scrubbing behind, in one go; the roller brush reaches up to 800 rpm with 20 kPa suction, and it supports 85°C hot-water scrubbing to handle grease in food-service settings, paired with an eight-in-one docking station for charging, water refilling, wastewater discharge, detergent refilling, and roller-brush cleaning/drying. Cleaning automation has mainly been deployed in large venues so far; the barrier to moving into small stores has always been that "the robot can get in, but still needs frequent human intervention."Source: AiThority source
Wulun Technology partners with Lens Technology's robotics unit on mass production of bionic humanoid heads · humanoid
The two companies signed a strategic cooperation framework agreement in Changsha on August 10, covering bionic robot busts and highly articulated head/face assemblies, jointly advancing mass production of bionic humanoid heads. Lens Robotics is a subsidiary of Lens Technology responsible for robot contract manufacturing and core smart-equipment production; the company says it has built annual production capacity for hundreds of thousands of robots, and this marks its zero-to-one expansion into the bionic-robotics segment. Source: China.com source
Hardware · Supply Chain
· Tactile sensors: Zhongke Guiji (Chinese tactile-sensor startup) places 200,000-unit order with Filna New Materials: the two companies signed a strategic cooperation agreement on August 12, under which Zhongke Guiji will procure 200,000 high-performance tactile sensors from Filna New Materials (605488.SH), which it says will support mass production of 40,000 industrial-grade dexterous hands. Zhongke Guiji was incubated from the State Key Laboratory of Multimodal Artificial Intelligence Systems at the Chinese Academy of Sciences' Institute of Automation; its Casia Hand series of dexterous hands has achieved 100% domestic sourcing of core components, with cumulative orders exceeding RMB 100 million in H1 2026. The company plans to reach stable annual production capacity of nearly 30,000 dexterous hands by the end of 2026 and exceed 100,000 units of annual capacity by the end of 2027 (⚠️ Company projection). Filna New Materials reported approximately RMB 2.8 billion in revenue in 2025, with its sensor business still in an early-stage phase. source
· The durability curse of dexterous hands: high-DOF tactile dexterous hands priced from the low hundred-thousands to several hundred thousand RMB often last only a few weeks to two or three months under continuous industrial operation — tendon-driven designs offer strong bio-mimicry and high DOF but lose force easily and are hard to maintain. Zhongke Huisi (Chinese dexterous-hand startup) chairman Zhang Zhengtao told National Business Daily that his solution is a "3-minute tendon swap," built into the company's pre-launch F-series tendon-driven dexterous hand, making tendon-replacement cost nearly negligible. GGII data: China's dexterous-hand market sold approximately 19,200 units in 2025, up 236.84% year-over-year, projected to reach 70,200 units in 2026 and potentially exceed 430,000 units by 2030; from 2023 through May 2026, China's dexterous-hand industry saw approximately 57 funding rounds totaling roughly RMB 12.8 billion. The company's CEO Yang Ping also noted that open-source fine-manipulation datasets for dexterous hands number only in the tens of thousands, orders of magnitude behind the millions of trajectories available for whole-robot manipulation data. source
Top comments (0)